Google Unveils Gemini agent, a Single Agent for Work - Claude Among the Model Options, Pricing and Launch Timing Not Disclosed

At Gemini at Work 2026 on October 8, Google Cloud announced Gemini agent. It keeps running in the cloud and can create coworker agents with their own Workspace accounts. Current models are the Gemini family and Claude. Google says SMBs have early access but gives no pricing, regions, or launch date.

Gemini Google-Cloud AI-agents Claude Google-Workspace enterprise-AI

Google Cloud introduced Gemini agent at its Gemini at Work 2026 event on October 8 (US time)¹. The official blog post, based on the keynote by Google Cloud CEO Thomas Kurian, frames it as a single, general-purpose agent for work. According to Google, one agent and one API cover answering questions, knowledge work, producing images and other media, and writing code and executing it.

For organizations weighing adoption, three points stand out: the design that gives an agent its own identity and account as a member of a team, the listing of Anthropic’s Claude next to the Gemini family as the models running underneath, and spending caps set per project. On the other hand, the announcement post gives no pricing, no regions, no eligible editions, and no launch date.

An agent you hand goals, not step-by-step instructions

Google describes an AI agent that receives an objective rather than instructions, takes on the outcome, and comes back with the work done¹. Besides assigning work directly, users can put tasks on a schedule or have the agent react to events. Google says it can serve one person as an assistant, act for a group the way a project manager would, or fill a defined role such as a finance analyst.

Execution happens in the cloud. Whichever device or channel a user comes in from, the agent keeps one set of memories and context, and jobs lasting hours or days carry on after the laptop lid is shut. The access points Google names are the web, iOS and Android, desktop apps for Mac and Windows, a command-line interface, Slack, Microsoft 365, and Google Workspace. Google adds that the agent can be embedded in third-party apps and can also run headless, without a dedicated interface.

For multi-step jobs, Gemini spins up temporary sub-agents on the fly and coordinates parallel and sequential steps over hours or days. Each sub-agent carries an identity of its own. TechCrunch reports that a “tasks inbox” view lets users follow Gemini’s reasoning, the hand-offs to sub-agents, which skills were loaded, the code, and overall progress³.

Examples of systems it can reach include Confluence, Teams, Slack, Git, Jira, Salesforce, ServiceNow, BigQuery, and Snowflake, and it can also talk to MCP servers on either side of the corporate network¹. Company-wide registries let teams and departments share the skills (reusable instructions, knowledge, and workflows) and tools they build.

Models today: the Gemini family and Claude, with other proprietary and open models later

On models, Google treats the agent and the underlying model as two separate choices¹. Each job goes to whichever model suits it. For now that means Gemini models and Anthropic’s Claude models; other leading proprietary and open models are slated for later, without any company or model names given. Google argues that fitting the model to the job improves accuracy on hard work and cuts the cost of easy work.

In its section on costs, the post says users can pick the right model for a job and, in larger projects, mix different models across specific tasks, but it does not describe where or how a user would specify one. TechCrunch reports that the AI chooses by default, that users can override the choice, and that third-party options begin with Claude³.

As a way to use Claude inside Workspace documents, Anthropic put Claude for Google Workspace, a sidebar for Docs, Sheets, and Slides, into beta on October 6. That is an Anthropic product used by signing in with a Claude account. This announcement is a different matter: Claude appears as one of the models that run inside Google’s agent.

In Workspace, the agent can join the team

Inside Workspace, Gemini operates directly in Gmail, Docs, Sheets, Slides, Drive, Calendar, and Chat, and keeps its memory, its skills, and the controls that apply on every other surface¹. Google groups its Workspace behavior into three modes.

The first is personal help. In Google’s example, a user asks for next week’s meeting with the regular group of regional event leads without naming anyone or giving addresses. Gemini identifies the people from the chat space membership and the thread about the previous event, looks at calendars, and opens an email thread to settle on a time, outside participants included. The second is proactive delegation: when a manager emails asking for a project update in slide form, Workspace Intelligence flags it as something that can be handed off and offers to hand it to Gemini with one click.

The third is the coworker agent. A user describes the role needed, Gemini creates the agent, and the agent is issued its own Workspace account: an email address (in the form @agents.company.com), a calendar, Drive, and a listing in the company directory. Colleagues bring it into a Chat space or @mention it, just as they would a person. Tagged in a Docs comment, it can propose an edit and answer in the comment thread, and version history records that activity under the agent’s own name.

Google says a coworker agent acts as itself, not as the person who asked, sees only what has been shared with it, and gets access through the sharing settings and memberships the team already has. With the Claude for Google Workspace add-on, by contrast, Claude’s edits are recorded in version history under the user’s own name. The two differ in whose name ends up on the record.

Constraining agents with identity, audit trails, and Agent Gateway

On governance, Google says each agent is given a cryptographically attested identity and managed like an employee under least privilege¹. That identity is stamped on its activity logs and on any VM started up to execute code for it. When the agent connects to an outside system, its identity is carried over through standards such as OAuth. Every action Gemini takes lands in an audit trail credited to the agent, not to a person.

All Gemini agents run inside Agent Sandbox, which has a network boundary of its own. All inbound, outbound, and agent-to-agent traffic is routed via Agent Gateway, which Google calls an AI network firewall. In Google’s example, a rule such as barring agents from opening documents labeled Need to Know is written once and then enforced across every agent in the company.

Hitting a spending cap pauses the project’s agent

On cost, Google lists Smart Routing and real-time spending caps in addition to matching models to jobs¹. Smart Routing, as Google describes it, sorts workloads automatically and sends each one to the model offering the best performance at the lowest cost.

Spending caps are hard limits on AI spend, configured for each project through Cloud Billing Console. Gemini tracks token consumption and sandbox costs, and once a cap is reached that project’s agent is paused. Work can be resumed from the console with a single click. Since costs are tracked project by project, Google says AI costs can be charged back to individual departments. When multi-day work is left running in the cloud, being able to define in advance when it stops is likely to be a basic operational requirement.

No pricing, regions, or launch date

The announcement post under Kurian’s name says nothing about how or when the agent will be offered¹. A separate blog post aimed at small and midsize companies, published the same day, says Gemini is now available to SMB customers through early access and will reach all customers soon². That post, too, gives no pricing, regions, or dates. TechCrunch reports that Google will roll the agent out to businesses first and to consumers afterward³.

For industry versions, the post says Financial Services and Legal are “in preview” and that Government, Healthcare, and Retail are “coming soon”¹. The financial and legal offerings are Gemini Enterprise for Financial Services and Gemini Enterprise for Legal, announced on August 25 and in preview from that point. This announcement does not move them to general availability.

How this relates to the industry editions and skills

Google has made a run of agent-related Gemini announcements over roughly the past two months. In August it added financial and legal editions to Gemini Enterprise that bundle industry-specific skills, connectors, and agents. On September 30 it announced that it was bringing reusable custom instructions called skills to the Gemini app and Workspace and phasing out Gems.

Gemini agent presents skills, tools, memory, and sub-agents together as one agent. The post does not explain how its company skills registry relates to the skills that arrived in Workspace.

For organizations already on Workspace or Gemini Enterprise, what can be checked today is the direction of the design. If agents hold dedicated accounts listed in the company directory, how will those accounts’ permissions and directory entries be managed? If work is routed across several models including Claude, how does data handling differ by model? These are questions to put to Google once pricing and terms are published. Because the agent is also reachable from Microsoft 365 and Slack, organizations that work mainly outside Workspace could also include it in their comparisons.